Hacker News Reader

Comment by modeless | Hacker News Reader

> There's nothing about a unified memory architecture that requires the GPU and CPU to be on the same die, let alone requires the RAM to be in the same package

No, but unified memory does require starving the GPU of bandwidth. There's a reason high end GPUs use different types of memory than CPUs.

Either Apple will be content with Mac Pro not having the absolute graphics performance lead, or they will make a discrete GPU with dedicated memory. I'm not sure which they will choose but I suspect the latter.

View on HN

The Xbox Series X (retail price $499) uses a 7nm SOC, where the GPU has 12 TFLOPS of single precision performance. The fastest single GPU Apple sells on the Mac Pro is 14 TFLOPS (with a step-up fee of $2,400!).

There doesn't seem to be any reason why Apple couldn't integrate a 16 TFLOPS GPU on an SOC using a 5nm or better process. The M1 GPU is 2.6 TFLOPS, they can x6 or more it, and obviously change the memory subsystem (HBM?).

Xbox Series X is very far from the top graphics available. GTX 3090 at 35 TFLOPS is what Apple needs to compete with for Mac Pro. Those chips are reticle limited which means there's simply no room to add a CPU, let alone 32 cores. You need separate dies.

Also, Xbox uses GDDR which trades off CPU performance for GPU performance. Apple probably wouldn't do that for a Mac Pro. The high end needs separate memory so the GPU can have bandwidth and the CPU can have low latency.

Well, depends. If they're looking to replace the low- and medium end of the Mac Pro offerings, a SOC would be fine. If they are looking to exceed it in every way, then a separate GPU with 16x the current M1 GPU cores is necessary. (The GPU on M1 takes up about 30mm2-40mm2). Put that on a TSMC N5P process, and that will be quite the setup...

Reticle limited on Samsung 8nm means plenty of room is available on TSMC 5nm.

Also, GDDR does not have appreciably higher latencies than DDR memory when measured in nanoseconds. It's just more expensive than DDR and much more limited in terms of capacity.

If Apple does go the "huge SoC" route I'd expect to see HBM2 memory with socketed DDR4 or DDR5. It'd provide the best of both worlds - extremely high bandwidth and low latency for a small portion (say 32-64GB) of the memory, and high capacity for the rest (say 1-2TB), all without compromising the unified memory concept.

This is not without precedent - recent Xeon Phis, for all their other shortcomings, have had a similar memory hierarchy.

Yeah but when discrete GPUs get to 5nm they will still be reticle limited. They're not going to stop getting bigger. If Apple maintains a process node advantage maybe they could do SoCs to compete with discrete graphics at the previous node but I think Nvidia will catch up.

I agree that a combination of HBM and DDR sounds pretty good for a unified memory architecture. Are you imagining it as just another layer in the cache hierarchy or something actively managed?

How is it a unified memory architecture if one part is faster than another? Sounds rather non-unified.

"Unified" in this case refers to the CPU and GPU having the same access to memory. This could still be true even if part of the memory is faster, as long as it's faster for both CPU and GPU.

I agree, i'd also expect more integration of the Afterburner tech with support for multiple codecs beyond the current version.

Reply on news.ycombinator.com